Large-scale dense low-altitude flight plan capacity balancing method
By constructing a spatiotemporal reachability network and reinforcement learning model, high-conflict, low-cost flight plans are identified and canceled, solving the problem of airspace capacity saturation in large-scale, dense low-altitude flights and achieving efficient and safe low-altitude flight management.
Patent Information
- Application Number
- CN202511742809.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-02-10
AI Technical Summary
In large-scale, dense low-altitude flights, traditional conflict management methods are ineffective in handling flight plan conflicts, leading to increased management complexity and airspace capacity saturation, which affects flight safety and efficiency.
By constructing a low-altitude flight plan conflict network based on spatiotemporal accessibility, a flight plan capacity balancing reinforcement learning model is designed. Using graph neural networks and SAC agent training methods, high-conflict, low-cost flight plans are identified and canceled to achieve airspace capacity balancing.
It effectively identifies and cancels critical conflict flight plans, reduces management difficulty, enhances airspace service capabilities, and ensures flight safety and efficiency, while possessing real-time and low-cost characteristics.
Smart Images
Figure CN121505930A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of aviation, and particularly relates to a large-scale dense low-altitude flight plan capacity balancing method. BACKGROUND
[0002] Low-altitude traffic has great application potential in providing convenient and high-speed transportation services and relieving ground traffic problems due to its three-dimensional space utilization characteristics. With the advent of various high-performance unmanned aerial vehicles and electric vertical take-off and landing aircraft, the service boundary of low-altitude traffic is constantly expanding, and multi-scenario integration, large-scale deployment and high-density operation have become an inevitable trend of future development. In the low-altitude airspace of high-density flight, the conflict of a single flight plan is easy to spread to the surrounding plans through the airspace association relationship, and the coupling effect between multi-conflicts will greatly increase the complexity of flight plan management. At the same time, when the number of flight requests exceeds the airspace carrying capacity, the effectiveness of the traditional flight conflict management method will decrease significantly, and it is difficult to realize the conflict resolution of all plans. In order to improve the safety level of low-altitude flight, the large-scale dense low-altitude flight plan is managed before flight, how to actively screen and reject high-conflict and low-value plans, so that the number of remaining low-altitude flight plans is close to or lower than the airspace capacity, thereby reducing the complexity of subsequent strategic conflict resolution, and finally realizing a low-altitude flight management scheme with high solving performance and strong service capability, which is of great significance to ensure safe and efficient operation of low-altitude flight. SUMMARY
[0003] Therefore, the application provides a large-scale dense low-altitude flight plan capacity balancing method to solve the problems in the background art.
[0004] To achieve the above-mentioned purpose, the application provides the following technical scheme:
[0005] Based on the time accessibility of the three-dimensional conflict area of each plan according to the airspace data and the maximum flight speed, a time-space conflict edge containing multi-dimensional indexes is designed, and a low-altitude flight plan conflict network based on time-space accessibility is constructed;
[0006] According to the application type priority and the purpose task value of each plan, the cancellation cost of each plan is calculated, and a flight plan capacity balancing reinforcement learning model is established in combination with the target of flight plan capacity balancing;
[0007] Based on the graph neural network, a key node recognition agent model is designed, and multi-scale feature capture of local influence and global situation is realized by introducing a virtual node and a node feature aggregation mechanism based on time-space conflict edges;
[0008] The artificial intelligent agent is trained by using an intelligent agent training method based on SAC (Soft Actor-Critic), so that a flight plan capacity balancing intelligent agent capable of effectively identifying high conflict and low cost flight plans is obtained.
[0009] The low-altitude flight plan capacity balancing scheme is obtained by identifying and canceling the dense low-altitude flight plan of the super airspace capacity by the intelligent agent, and reducing the flight plan demand under the airspace capacity constraint.
[0010] Optionally, the low-altitude flight plan conflict network based on spatiotemporal reachability comprises:
[0011] Each low-altitude flight plan is compared with each other, and it is determined whether there is a flight plan conflict between the two based on the spatiotemporal reachability principle;
[0012] The flight plan conflict comprises:
[0013] (1) Height conflict determination: height conflict focuses on the overlap relationship of the vertical airspace range of the two plans. If there is no overlap, it is considered that there is no height conflict, otherwise it is determined that there is a height conflict.
[0014] (2) Two-dimensional airspace conflict determination: two-dimensional airspace conflict pays attention to whether the safety interval of the horizontal airspace (longitude and latitude range) of the two plans meets the requirements, and defines as the set of all points in the horizontal airspace of flight plan a. If there exists a point and a point whose Euclidean distance is less than the safety interval , it is determined that there is a two-dimensional airspace conflict.
[0015] (3) Time conflict determination: time conflict needs to be determined in combination with the spatiotemporal reachability of the two-dimensional airspace conflict points. For points and in the airspace of the two plans that do not meet the safety interval requirement, the earliest time and the latest time of the two plans to reach the corresponding points are calculated. If the time window overlaps, it is determined that there is a time conflict.
[0016] (4) Comprehensive conflict determination: if the two flight plans simultaneously exist height conflict, two-dimensional airspace conflict and time conflict, it is considered that there is a conflict between the two plans.
[0017] For the flight plans that exist in space-time, a directed space-time conflict edge containing multi-dimensional indexes is established between them, reflecting the influence of one plan on the space-time conflict of another plan. The space-time conflict edge comprises:
[0018] (1) Conflict time proportion: the calculation method is the ratio of the time with conflict between the flight plans to the total time length of the flight plans.
[0019] (2) Conflict area ratio: The calculation method is the ratio of the two-dimensional airspace area / route length where there is a conflict between flight plans to the total airspace area / route length of the flight plan applied for.
[0020] Optionally, the flight plan capacity balancing reinforcement learning model includes:
[0021] Taking into account the application type, priority, objective value, and service capabilities of low-altitude flight plans, the cost of canceling such plans can be calculated as follows:
[0022]
[0023] in, The relative priority of the plan type and These are the economic value coefficient and social value coefficient of the planned objectives and tasks, respectively. and These are the weighting coefficients. The service capacity of the plan is reflected in the total flight distance of the route for route-type plans and the perimeter of the polygonal airspace for airspace-type plans.
[0024] The objective of establishing a low-altitude flight plan capacity balance model is to minimize the number of canceled flight plans as the first objective function and to minimize the total cost of canceling flight plans as the second objective function.
[0025] The first objective function is:
[0026]
[0027] in, The number of canceled flight schedules;
[0028] The second objective function is:
[0029]
[0030] in, The cost of canceling the flight plan;
[0031] A reinforcement learning model is established for the corresponding objective. The low-altitude flight plan capacity balancing problem is modeled as selecting and removing a node (canceling the corresponding plan) in each round of interaction between the agent and the environment until the termination condition is met. The termination condition can be expressed as:
[0032]
[0033] in, For subgraph The number of nodes, The set consisting of all subgraphs. The maximum acceptable subgraph size can be set based on the processing capacity of subsequent strategic conflict management algorithms; Let the set of weights of the remaining edges be denoted as . The maximum acceptable edge weight reflects the most severe spatiotemporal conflict that can be tolerated;
[0034] Considering both the extent of network disruption and the cost of plan cancellation, a composite reward function is designed, including:
[0035] (1) End-of-scene reward: This reward is applied at the end of each scene during training and can be represented as:
[0036]
[0037] Where t is the time step when done=1. and These represent the time steps when done=1 under the predicted poor and good strategies, respectively, taking 2 / 3 and 1 / 3 of the total number of nodes. The scaling factor for the final reward.
[0038] (2) Single-step reward: used to provide real-time feedback on the quality of the decision, including:
[0039] 1) Large Subgraph Reduced Reward: This reward function aims to encourage the agent to prioritize processing the largest subgraph with the most complex conflict propagation impact in each decision, and can be represented as:
[0040]
[0041] in, The size of the largest subgraph at time step t. Let t be the set of all subgraphs at time step t; The ranking is based on the size of the subgraph; the larger the subgraph, the higher its ranking. Reduce the scaling factor of rewards for large subgraphs;
[0042] 2) Subgraph Reduction Reward: This reward focuses on the destructive effect of removing a node on its subgraph, guiding the agent to choose nodes that can split the subgraph into smaller pieces. It can be represented as:
[0043]
[0044] in, For the agent to select the subgraph to which the rejection flight plan belongs at time step t-1, The superlinear coefficients are used to provide a more significant reward for large block splits. Reduce the scaling factor of the reward for the subgraph;
[0045] 3) Spatiotemporal Conflict Mitigation Reward: This reward quantifies the direct conflict mitigation effect of removing nodes, guiding the agent to prioritize nodes with significant spatiotemporal overlap with other plans. It can be represented as:
[0046]
[0047] in, Let t be the set of remaining unrejected low-altitude flight plans. To choose the flight plan to refuse, A scaling factor to mitigate rewards for spatiotemporal conflicts.
[0048] 4) Flight plan cancellation penalty: This penalty is used to constrain the agent's decision-making risk and prevent the agent from rejecting high-priority, high-value plans in pursuit of high rewards. It can be represented as:
[0049]
[0050] in, The cost of canceling flight plan A The scaling factor for refusing penalties to flight plans.
[0051] Optionally, the key node identification agent model includes:
[0052] Virtual nodes are introduced to connect with all other flight plan nodes, providing global situational information when node information is aggregated;
[0053] By mapping node and edge features to the same dimension using a node encoder and an edge encoder, we obtain the node embedding. and edge embedding ;
[0054] Based on the edge connections in the graph, construct a triplet structure of "source node-directed edge-target node" to extract the source node embedding. Embedding the target node ;
[0055] Using a multi-head graph attention mechanism that incorporates edge features, the attention score of the source node can be expressed as:
[0056]
[0057] in, , , , , , These are trainable parameters. It is a learnable gating function;
[0058] The association information passed from the source node to the target node based on edge features can be represented as:
[0059]
[0060] in, , These are trainable parameters;
[0061] Aggregating all attention messages passed from the source nodes to the target node, the embedding of the target node can be represented as:
[0062]
[0063] in, Let be the set of neighboring nodes of node i. For the attention information of the directed edge from j to i, The attention score is given for the directed edge from j to i. It is a small bias term to avoid a denominator of 0.
[0064] The original embedding of the target node through the gating layer Weight scaling is performed, and the attention aggregation results are concatenated to generate the final embedding of the target node. ;
[0065] The node features are mapped to 1D after passing through an FNN layer and then normalized using softmax to obtain the removal preference for each node.
[0066] Optionally, the SAC-based agent training method includes:
[0067] The initial training settings construct the neural network to be trained, including one policy network, which selects a key node to identify the agent model; two value networks, which are consistent with the feature extraction part of the policy network, but after attention aggregation, they only output the Q value estimate of each action through a fully connected layer and do not perform softmax operation; and two target value networks, which correspond one-to-one with the value networks, and the parameters are synchronized from the value networks through a soft update mechanism.
[0068] The experience gained by the agent interacting with the environment is stored in a replay buffer. When the replay buffer reaches a specified size, the experience stored in the replay buffer is updated after each interaction. Combinations are extracted in batches, and the policy network, value network, and... The loss is calculated and updated based on the Adam optimizer. The three loss functions can be expressed as:
[0069]
[0070]
[0071]
[0072] in, For the parameters of the policy network, For the parameters of value network i, For the parameters of the target value network i, The target entropy;
[0073] SAC's target network uses a soft update method, which can be represented as:
[0074]
[0075] in, It is located between 0 and 1 to ensure that some information of the original network is still retained with each update, thus improving training stability.
[0076] Optionally, the low-altitude flight plan capacity balancing scheme method includes:
[0077] Construct a conflict network environment based on the conflict relationships between various flight plans;
[0078] The trained agent identifies critical conflict flight plans based on the current network state and cancels them, removes the corresponding node of the plan in the flight plan network, and updates the network.
[0079] Repeat the above steps until the termination conditions are met, thus completing the low-altitude flight plan capacity balance.
[0080] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0081] This invention provides a capacity balancing method for large-scale dense low-altitude flight plans. Addressing differentiated low-altitude flight plan demands, it utilizes reinforcement learning and graph neural networks to identify and cancel critical conflicting flight plans, effectively decoupling conflict relationships between flight plans, reducing flight plan demands to within airspace capacity limits, and lowering the difficulty of flight plan management. Compared to existing methods that identify key network nodes based on network centrality indicators, the method described in this invention removes fewer nodes at a lower cost and achieves near real-time solution efficiency. It is highly practical, effectively ensuring flight safety, reducing management complexity, and enhancing airspace service capabilities.
[0082] This invention can effectively solve the problem of airspace capacity saturation and difficulty in management caused by the large-scale and dense flight demand of drones in the future. It takes into account both conflict mitigation capabilities and plan execution assurance, accurately identifies key conflict flight plans, and balances plan demand and airspace capacity. It is of great significance for improving airspace service capabilities in dense flight scenarios. Attached Figure Description
[0083] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:
[0084] Figure 1 This is a flowchart illustrating the capacity balancing method for large-scale intensive low-altitude flight plans of the present invention.
[0085] Figure 2 This is a schematic diagram of the intelligent agent model structure for key node identification in this invention;
[0086] Figure 3 This is an embodiment of the invention showing the change in the size of the conflict subgraph in the low-altitude flight plan capacity balancing process.
[0087] Figure 4 This is a distribution diagram of the cost of canceling low-altitude flight plan capacity balancing provided in one embodiment of the present invention. Detailed Implementation
[0088] To make the objectives, uses, and technical solutions of this invention easier to understand, and to provide a clearer and more complete description of the invention, the following detailed description is provided in conjunction with embodiments. Obviously, the described embodiments are only some embodiments of this invention, and it should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the scope of protection of this invention.
[0089] The application principle of the present invention will be described in detail below with reference to the accompanying drawings.
[0090] like Figure 1 As shown in Example 1: This example provides a method for balancing the capacity of large-scale, dense low-altitude flight plans, including:
[0091] Step A: Based on the airspace data and maximum flight speed of each plan, the temporal reachability of its three-dimensional conflict area is deduced. A spatiotemporal conflict edge containing multi-dimensional indicators is designed, and a low-altitude flight plan conflict network based on spatiotemporal reachability is constructed, such as... Figure 2 As shown, the specific steps include:
[0092] Step A1: Compare each low-altitude flight plan pairwise, and determine whether there is a conflict between the two flight plans based on the principle of spatiotemporal reachability.
[0093] (1) Determination of altitude conflict: Altitude conflict focuses on the overlap of the vertical airspace range of the two plans. If there is no overlap, it is considered that there is no altitude conflict; otherwise, it is determined that there is an altitude conflict.
[0094] (2) Two-dimensional airspace conflict determination: Two-dimensional airspace conflict focuses on whether the safety interval between the two planned horizontal airspaces (latitude and longitude range) meets the requirements, and is defined as follows: Given the set of all points in the horizontal airspace of flight plan a, if there exists a point With point The Euclidean distance is less than the safety interval Then it is determined that there is a two-dimensional spatial conflict;
[0095] (3) Time conflict determination: Time conflicts need to be determined in conjunction with the spatiotemporal accessibility of conflicting points in the two-dimensional airspace. For points in the two planned airspaces that do not meet the safety interval requirements, and Calculate the earliest and latest times for the two plans to arrive at the corresponding point. If the time windows overlap, it is determined that there is a time conflict.
[0096] (4) Comprehensive conflict determination: If two flight plans have conflicts in altitude, two-dimensional airspace, and time, then the two plans are considered to be in conflict.
[0097] Step A2: For flight plans involving spatiotemporal conflict, establish directed spatiotemporal conflict edges containing multidimensional indicators to reflect the spatiotemporal conflict affecting one plan with another. The spatiotemporal conflict edges include:
[0098] (1) Conflict time percentage: The calculation method is the ratio of the time of conflict between flight plans to the total application time;
[0099] (2) Conflict area ratio: The calculation method is the ratio of the two-dimensional airspace area / route length where there is a conflict between flight plans to the total airspace area / route length of the flight plan applied for.
[0100] Step B: Based on the application type priority and objective mission value of each plan, the cancellation cost of each plan is calculated. Combined with the goal of flight plan capacity balancing, a reinforcement learning model for flight plan capacity balancing is established, specifically including:
[0101] Step B1: Taking into account the application type priority, objective mission value, and service capabilities of the low-altitude flight plan, calculate the cost of plan cancellation, which can be expressed as:
[0102]
[0103] in, The relative priority of the plan type and These are the economic value coefficient and social value coefficient of the planned objectives and tasks, respectively. and These are the weighting coefficients. The service capacity of the plan is reflected in the total flight distance of the route for route-type plans and the perimeter of the polygonal airspace for airspace-type plans.
[0104] Step B2: The objective of establishing a low-altitude flight plan capacity balance model is to minimize the number of canceled flight plans as the first objective function and to minimize the total cost of canceling flight plans as the second objective function.
[0105] The first objective function is:
[0106]
[0107] in, The number of canceled flight schedules;
[0108] The second objective function is:
[0109]
[0110] in, The cost of canceling the flight plan;
[0111] Step B3: Establish a reinforcement learning model corresponding to the objective, modeling the low-altitude flight plan capacity balancing problem as selecting and removing a node (canceling the corresponding plan) in each round of interaction between the agent and the environment, until the termination condition is met. The change in the size of the conflict subgraph during the removal process is as follows: Figure 3 As shown, the termination condition can be expressed as:
[0112]
[0113] in, For subgraph The number of nodes, The set consisting of all subgraphs. The maximum acceptable subgraph size can be set based on the processing capacity of subsequent strategic conflict management algorithms; Let the set of weights of the remaining edges be denoted as . The maximum acceptable edge weight reflects the most severe spatiotemporal conflict that can be tolerated;
[0114] Step B4: Considering both the extent of network disruption and the cost of canceling the plan, design a composite reward function, including:
[0115] (1) End-of-scene reward: This reward is applied at the end of each scene during training and can be represented as:
[0116]
[0117] Where t is the time step when done=1. and These represent the time steps when done=1 under the predicted poor and good strategies, respectively, taking 2 / 3 and 1 / 3 of the total number of nodes. The scaling factor for the final reward.
[0118] (2) Single-step reward: used to provide real-time feedback on the quality of the decision, including:
[0119] 1) Large Subgraph Reduced Reward: This reward function aims to encourage the agent to prioritize processing the largest subgraph with the most complex conflict propagation impact in each decision, and can be represented as:
[0120]
[0121] in, The size of the largest subgraph at time step t. Let t be the set of all subgraphs at time step t; The ranking is based on the size of the subgraph; the larger the subgraph, the higher its ranking. Reduce the scaling factor of rewards for large subgraphs;
[0122] 2) Subgraph Reduction Reward: This reward focuses on the destructive effect of removing a node on its subgraph, guiding the agent to choose nodes that can split the subgraph into smaller pieces. It can be represented as:
[0123]
[0124] in, For the agent to select the subgraph to which the rejection flight plan belongs at time step t-1, The superlinear coefficients are used to provide a more significant reward for large block splits. Reduce the scaling factor of the reward for the subgraph;
[0125] 3) Spatiotemporal Conflict Mitigation Reward: This reward quantifies the direct conflict mitigation effect of removing nodes, guiding the agent to prioritize nodes with significant spatiotemporal overlap with other plans. It can be represented as:
[0126]
[0127] in, Let t be the set of remaining unrejected low-altitude flight plans. To choose the flight plan to refuse, A scaling factor to mitigate rewards for spatiotemporal conflicts.
[0128] 4) Flight plan cancellation penalty: This penalty is used to constrain the agent's decision-making risk and prevent the agent from rejecting high-priority, high-value plans in pursuit of high rewards. It can be represented as:
[0129]
[0130] in, The cost of canceling flight plan A The scaling factor for refusing penalties to flight plans.
[0131] Step C: An artificial intelligence agent was designed based on a graph neural network to identify key conflict nodes. By introducing virtual nodes and a node feature aggregation mechanism based on spatiotemporal conflict edges, multi-scale feature capture of local influences and global situational awareness was achieved. Figure 4 As shown, it specifically includes:
[0132] Step C1: Introduce virtual nodes to connect with all other flight plan nodes to provide global situational information when node information is aggregated;
[0133] Step C2: Map node and edge features to the same dimension using a node encoder and an edge encoder to obtain the node embedding. and edge embedding ;
[0134] Step C3: Based on the edge connections in the graph, construct a triplet structure of "source node-directed edge-target node" to extract the source node embedding. Embedding the target node ;
[0135] Step C4: Employ a multi-head graph attention mechanism that incorporates edge features. The attention score of the source node can be expressed as:
[0136]
[0137] in, , , , , , These are trainable parameters. It is a learnable gating function;
[0138] The association information passed from the source node to the target node based on edge features can be represented as:
[0139]
[0140] in, , These are trainable parameters;
[0141] Aggregating all attention messages passed from the source nodes to the target node, the embedding of the target node can be represented as:
[0142]
[0143] in, Let be the set of neighboring nodes of node i. For the attention information of the directed edge from j to i, The attention score is given for the directed edge from j to i. It is a small bias term to avoid a denominator of 0.
[0144] Step C5: Initial embedding of the target node through the gating layer Weight scaling is performed, and the attention aggregation results are concatenated to generate the final embedding of the target node. ;
[0145] Step C6: Map the node features to 1D after passing them through an FNN layer and perform softmax normalization to obtain the removal preference for each node.
[0146] Step D: Train the AI agent using the SAC (Soft Actor-Critic) algorithm to obtain a flight plan capacity balancing agent capable of effectively identifying high-conflict, low-cost flight plans. Specifically, this includes:
[0147] Step D1: Initialize training settings. Construct the neural network to be trained, including 1 policy network, 2 value networks, and 2 target value networks;
[0148] Step D2: The agent interacts with the environment and stores the learned experiences in the replay buffer. Once the replay buffer reaches a specified size, the experiences stored in the buffer are added after each interaction. Combinations are extracted in batches, and the policy network, value network, and... The loss is calculated and updated based on the Adam optimizer. The three loss functions can be expressed as:
[0149]
[0150]
[0151]
[0152] in, For the parameters of the policy network, For the parameters of value network i, For the parameters of the target value network i, The target entropy;
[0153] Step D3: The target network of SAC adopts a soft update method, which can be represented as:
[0154]
[0155] in, It is located between 0 and 1 to ensure that some information of the original network is still retained with each update, thus improving training stability.
Claims
1. A capacity balancing method for large-scale, intensive low-altitude flight plans, characterized in that, Includes the following steps: Based on the airspace data and maximum flight speed of each plan, the time accessibility of its three-dimensional conflict area is deduced, a spatiotemporal conflict edge containing multi-dimensional indicators is designed, and a low-altitude flight plan conflict network based on spatiotemporal accessibility is constructed. The cancellation cost of each plan is calculated based on the application type priority and objective mission value. Combined with the goal of flight plan capacity balancing, a reinforcement learning model for flight plan capacity balancing is established. A key node identification agent model is designed based on graph neural networks. By introducing virtual nodes and a node feature aggregation mechanism based on spatiotemporal conflict edges, multi-scale feature capture of local influence and global situation is achieved. The AI agent is trained using the SAC-based agent training method to obtain a flight plan capacity balancing agent that can effectively identify high-conflict, low-cost flight plans. By relying on intelligent agents to identify and cancel dense low-altitude flight plans that exceed airspace capacity, the flight plan demand is reduced to within airspace capacity constraints, resulting in a low-altitude flight plan capacity balance scheme.
2. The method for balancing the capacity of large-scale intensive low-altitude flight plans according to claim 1, characterized in that, The spatiotemporal reach-based low-altitude flight plan conflict network includes: By comparing each low-altitude flight plan in pairs, and based on the principle of spatiotemporal accessibility, it is determined whether there is a conflict between the two flight plans; For flight plans that exist in space and time, a directed space-time conflict edge containing multi-dimensional indicators is established to reflect the impact of space-time conflict between one plan and another plan.
3. The capacity balancing method for large-scale intensive low-altitude flight plans according to claim 1, characterized in that, The flight plan capacity balancing reinforcement learning model includes: Taking into account the application type, priority, objective mission value, and service capabilities of low-altitude flight plans, calculate the cost of plan cancellation. The objective is to establish a capacity balance model for low-altitude flight plans. A reinforcement learning model is established to address the target problem. The capacity balancing problem of low-altitude flight plans is modeled as selecting and removing a node in each round of interaction between the agent and the environment until a termination condition is met. The termination condition is expressed as: ； in, For subgraph The number of nodes, The set consisting of all subgraphs. The maximum acceptable subgraph size is set based on the processing capacity of the subsequent strategic conflict management algorithm; Let the set of weights for the remaining edges be denoted as . The maximum acceptable edge weight reflects the most severe spatiotemporal conflict that can be tolerated; A composite reward function is designed, taking into account both the extent of network disruption and the cost of canceling the plan.
4. The capacity balancing method for large-scale intensive low-altitude flight plans according to claim 1, characterized in that, The key node identification intelligent agent model includes: Virtual nodes are introduced to connect with all other flight plan nodes, providing global situational information when node information is aggregated; By mapping node and edge features to the same dimension using a node encoder and an edge encoder, we obtain the node embedding. and edge embedding ; Based on the edge connections in the graph, construct a triplet structure of "source node-directed edge-target node" to extract the source node embedding. Embedding the target node ; A multi-head graph attention mechanism incorporating edge features is employed to aggregate source node features and directed edge features into the target node, and the original embedding of the target node is then processed through a gating layer. Weight scaling is performed, and the attention aggregation results are concatenated to generate the final embedding of the target node. ; The node features are mapped to 1D after passing through an FNN layer and then normalized using softmax to obtain the removal preference for each node.
5. The capacity balancing method for large-scale intensive low-altitude flight plans according to claim 1, characterized in that, The SAC-based agent training method includes: Initialize training settings and build the neural network to be trained; The experience gained by the agent interacting with the environment is stored in a replay buffer. When the replay buffer reaches a specified size, the experience stored in the replay buffer is updated after each interaction. Combinations are extracted in batches, and the policy network, value network, and... The loss is calculated and updated based on the Adam optimizer. The three loss functions are expressed as follows: ； ； ； in, For the parameters of the policy network, For the parameters of value network i, For the parameters of the target value network i, The target entropy; SAC's target network uses a soft update method, represented as follows: ； in, It is located between 0 and 1 to ensure that some information of the original network is still retained with each update, thus improving training stability.
6. The method for balancing the capacity of large-scale intensive low-altitude flight plans according to claim 1, characterized in that, The low-altitude flight plan capacity balancing scheme method includes: Construct a conflict network environment based on the conflict relationships between various flight plans; The trained agent identifies critical conflict flight plans based on the current network state and cancels them, removes the corresponding node of the plan in the flight plan network, and updates the network. Repeat the above steps until the termination conditions are met, thus completing the low-altitude flight plan capacity balance.
7. The capacity balancing method for large-scale intensive low-altitude flight plans according to claim 2, characterized in that, The flight plan conflicts include: Altitude conflict determination: Altitude conflict focuses on the overlap of the vertical airspace of the two plans. If there is no overlap, there is no altitude conflict; otherwise, an altitude conflict is determined. Two-dimensional airspace conflict determination: Two-dimensional airspace conflict focuses on whether the safety separation between two planned horizontal airspaces meets the requirements, defined as follows: Given the set of all points in the horizontal airspace of flight plan a, if there exists a point With point $ The Euclidean distance is less than the safety interval Then it is determined that there is a two-dimensional spatial conflict; Time conflict determination: Time conflicts need to be determined in conjunction with the spatiotemporal reachability of conflicting points in two-dimensional airspace. For points in the two planned airspaces that do not meet the safety interval requirements... and Calculate the earliest and latest times for the two plans to arrive at the corresponding point. If the time windows overlap, it is determined that there is a time conflict. Comprehensive conflict determination: If two flight plans simultaneously have conflicts in altitude, two-dimensional airspace, and time, then the two plans are considered to be in conflict. The spatiotemporal conflict edges include: Conflict time percentage: Calculated as the ratio of the time of conflict between flight plans to the total time of their application; Conflict area ratio: The calculation method is the ratio of the area of two-dimensional airspace that conflicts between flight plans to the length of the flight route, and the total area of the airspace applied for to the total length of the flight route.
8. A method for balancing the capacity of large-scale, intensive low-altitude flight plans according to claim 3, characterized in that, The cost of canceling the plan is expressed as: ； in, The relative priority of the plan type and These are the economic value coefficient and social value coefficient of the planned objectives and tasks, respectively. and These are the weighting coefficients. The service capacity of the plan is reflected in the total flight distance of the route for route-type plans and the perimeter of the polygonal airspace for airspace-type plans. The low-altitude flight plan capacity balancing takes minimizing the number of canceled flight plans as the first objective function and minimizing the total cost of canceling flight plans as the second objective function. The first objective function is: ； in, The number of canceled flight schedules; The second objective function is: ； in, The cost of canceling the flight plan; The composite reward function includes: End-of-scene bonus: This bonus is applied at the end of each scene during training and is represented as: ； Where t is the time step when done=1. and These represent the time steps when done=1 under the predicted poor and good strategies, respectively, taking 2 / 3 and 1 / 3 of the total number of nodes. The scaling factor for the final reward; Single-step reward: Used to provide real-time feedback on the quality of the decision, including: (1) Large subgraphs reduce rewards: This reward function aims to encourage the agent to prioritize processing the largest subgraph with the most complex conflict propagation effects in each decision, which can be shown as: ； in, The size of the largest subgraph at time step t. Let t be the set of all subgraphs at time step t; The ranking is based on the size of the subgraph; the larger the subgraph, the higher its ranking. Reduce the scaling factor of rewards for large subgraphs; (2) Subgraph Reduction Reward: This reward focuses on the destructive effect of removing a node on its subgraph, guiding the agent to select nodes that can split the subgraph into smaller pieces, as shown below: ； in, For the agent to select the subgraph to which the rejection flight plan belongs at time step t-1, The superlinear coefficients are used to provide a more significant reward for large block splits. Reduce the scaling factor of the reward for the subgraph; (3) Spatiotemporal conflict mitigation reward: This reward quantifies the direct conflict mitigation effect of removing nodes, guiding the agent to prioritize nodes with severe spatiotemporal overlap with other plans, expressed as: ； in, Let t be the set of remaining unrejected low-altitude flight plans. To choose the flight plan to refuse, A scaling factor to mitigate rewards for spatiotemporal conflicts; (4) Flight plan cancellation penalty: This penalty is used to constrain the agent's decision-making risk and prevent the agent from rejecting high-priority, high-value plans in pursuit of high rewards. It can be expressed as: ； in, The cost of canceling flight plan A Scaling factor for refusing penalties on flight plans.
9. A capacity balancing method for large-scale intensive low-altitude flight plans according to claim 4, characterized in that, The multi-head graph attention mechanism that incorporates edge features includes: The attention score of the source node is represented as: ； in, , , , , , These are trainable parameters. It is a learnable gating function; The association information passed from the source node to the target node based on edge features is represented as follows: ； in, , These are trainable parameters; Aggregate the attention messages passed from all source nodes to the target node to obtain the embedding representation of the target node as follows: ； in, Let be the set of neighboring nodes of node i. For the attention information of the directed edge from j to i, The attention score is given for the directed edge from j to i. It is a small bias term to avoid a denominator of 0.
10. A capacity balancing method for large-scale intensive low-altitude flight plans according to claim 5, characterized in that, The construction of the neural network to be trained includes: One policy network is used, and key nodes are selected to identify the agent model; Two value networks are used, which are consistent with the feature extraction part of the policy network, but after attention aggregation, only the Q-value estimate of each action is output through a fully connected layer, without performing a softmax operation; There are two target value networks, each corresponding to a value network, and the parameters are synchronized from the value network through a soft update mechanism.